Projects

Describe the place you want in plain English; get five real ones back, each with a reason. An LLM parses your query, a daily research pass grounds it in real places, and a second LLM pass explains why each one fits.

Treating LLMs as unreliable tools instead of oracles. Custom MCP servers expose Gemini and GPT as tools inside Claude Code; a cross-vendor "Gauntlet" runs code and designs through an adversarial review panel (blind critique → crossfire rebuttal → verified verdict); fallback chains degrade instead of failing; and nightly launchd agents maintain my notes idempotently.

A personal fork of VoiceInk, the open-source macOS dictation app — push-to-talk anywhere, transcribed on-device by Whisper or Parakeet, cleaned up by an LLM before it lands at your cursor. The fork adds a voice engine: a weekly pipeline that distills my writing voice from my own dictation corpus into eval-gated style profiles (champion/challenger promotion, versioned rollback) and injects them into the enhancement prompt, plus nightly vocabulary mining from my Obsidian vault so transcription stops mangling domain terms. Recently hardened with a 31-finding multi-agent audit: anti-aliased resampling, real RIFF parsing, actor-isolated model lifecycle, and failure paths that speak instead of silently dropping dictations.

A daily map game. Five places, one tap each — drop your pin on a label-free satellite globe and get scored by how close you land, out of 1,000. Every place carries a two-sentence story of why it matters; streaks carry a shield. Ninety days of fact-checked puzzles ship in the bundle, so the whole game runs offline — same puzzle for everyone at local midnight, on iOS and the web.

A strength log that thinks like a coach. Every set is one tap — the app pre-fills weight and reps, then explains its next suggestion in plain English: +5 lb because you hit 8 reps on all sets. Progression reads your reps-in-reserve, deloads trigger off your trend instead of the calendar, and plate math is drawn to scale in competition colors. Local-first: your log lives on your phone, exports whenever you want, and works with no signal in a basement gym.

Real-time powder conditions for 53 ski resorts. Pulls snowpack, forecasts, road conditions, and webcams into a single score, with trip planning to sort out who's driving. Web and iOS.

A spatial-transcriptomics pipeline for Vizgen MERSCOPE mouse brain. QC against bulk RNAseq, Scanpy clustering, cell-type mapping with a per-cell confidence score, and registration of each section to the Allen CCFv3 atlas with STalign. Covered by an offline pytest suite and CI.

A Stanford Python library for multiplexed imaging analysis, covering the full pipeline from cell segmentation to spatial analysis. Introduces Patch Proximity Analysis. Published on PyPI and bioRxiv.